HySparse: A Hybrid Sparse Attention Architecture with Oracle Token Selection and KV Cache Sharing

TL;DR

HySparse introduces a hybrid sparse attention architecture that uses full attention layers as an oracle for token importance and reuses KV caches, significantly improving efficiency and performance in large models.

Contribution

HySparse presents a novel hybrid sparse attention architecture that leverages full attention as an oracle and shares KV caches, addressing key limitations of prior sparse attention methods.

Findings

Outperforms full attention and baseline models across various settings.

Reduces KV cache storage by nearly 10x in large models.

Achieves substantial performance gains with fewer full attention layers.

Abstract

This work introduces Hybrid Sparse Attention (HySparse), a new architecture that interleaves each full attention layer with several sparse attention layers. While conceptually simple, HySparse strategically derives each sparse layer's token selection and KV caches directly from the preceding full attention layer. This architecture resolves two fundamental limitations of prior sparse attention methods. First, conventional approaches typically rely on additional proxies to predict token importance, introducing extra complexity and potentially suboptimal performance. In contrast, HySparse uses the full attention layer as a precise oracle to identify important tokens. Second, existing sparse attention designs often reduce computation without saving KV cache. HySparse enables sparse attention layers to reuse the full attention KV cache, thereby reducing both computation and memory. We evaluate HySparse on both 7B dense and 80B MoE models. Across all settings, HySparse consistently outperforms both full attention and hybrid SWA baselines. Notably, in the 80B MoE model with 49 total layers, only 5 layers employ full attention, yet HySparse achieves substantial performance gains while reducing KV cache storage by nearly 10x.